Reddit user exports 35 streaming Claude thinking summaries; full chain-of-thought stays hidden

Additional-Ladder-24 · reddit · 2026-10-26

A Reddit user ran a controlled test on Claude Sonnet at Extra effort and captured 1 thinking block, 35 "thinkingsummarydelta" events streamed before the final answer — all 35 summaries recoverable from conversation history — while the full internal reasoning remained empty and marked hidden. He stresses he did not access hidden chain-of-thought, only confirmed user-visible reasoning summaries are transmitted and preservable.

He is building three components: a conversation exporter, a persistent memory/retrieval engine, and an independent RAG evaluation toolkit. The envisioned product is an audit-case memory system for AI agencies, storing past investigations, failed hypotheses, retrieval rankings, metrics, tool executions, and confirmed root causes so an assistant can search across cases and suggest diagnostic steps.

He asks practitioners five questions: whether debugging/auditing favors full reasoning traces, concise summaries, or structured execution evidence; whether thinking traces ever revealed problems logs and metrics couldn't; whether searchable investigation memory saves time; where the capability fits best (RAG audits, coding agents, autonomous workflows, incident investigation); and whether anyone has paid for such a solution. The goal is validating the problem before investing more development time.

Original post →

More from coding & agent

coding & agent channel →